fix(model): declare real WAV sizes in the final streamed audio block - #3242

AgentScope Version

2.0.10dev (main @ e4c1e07)

Description

Summary

In streaming mode, the final is_last audio DataBlock kept the streaming WAV header's 0xFFFFFFFF RIFF/data placeholder sizes instead of the real chunk sizes. The completed audio block is now finalized into a well-formed WAV, matching what the non-streaming path produces for the same call.

Why

• _build_streaming_wav_header() (src/agentscope/_utils/_audio.py:47-73) deliberately sets the RIFF and data chunk sizes to 0xFFFFFFFF since the total length isn't known yet, and its docstring states: "For a complete buffer use _build_wav_header instead, which declares real chunk sizes" (src/agentscope/_utils/_audio.py:64-65).
• Both streaming parsers prefix only the first audio delta with that header (src/agentscope/model/_openai_chat/_model.py:412, src/agentscope/model/_dashscope/_model.py:412). The stream accumulator then joins the deltas verbatim into the final is_last block via _AccBase64Source.build(), so the completed artifact kept the placeholder sizes.
• The non-streaming path already wraps the payload in a real header (src/agentscope/model/_openai_chat/_model.py:519, fix(model): wrap non-streaming chat audio in a real WAV header #3028), so streaming and non-streaming responses disagreed: consumers of the final block that trust the header — wave.getnframes(), duration, seeking, strict players — saw 4294967295 instead of the actual audio length.

Changes

• src/agentscope/_utils/_audio.py: add _finalize_streaming_wav(), which rewrites bytes 4-8 and 40-44 of a complete streaming buffer with the real RIFF (36 + data_len) and data (data_len) sizes. Buffers without the placeholder header are returned unchanged, so complete WAVs and other audio payloads are untouched.
• src/agentscope/model/_utils.py: _AccBase64Source.build() finalizes the joined buffer when media_type == "audio/wav". This is the single place where streamed audio deltas are joined for the final response, so both the OpenAI Chat and DashScope stream parsers are covered.
• The live per-delta chunks are unchanged — only the final accumulated block is affected.

Testing

• New regression tests TestOpenAIChatStream::test_stream_audio_final_block_declares_real_wav_sizes (tests/model_openai_chat_test.py) and TestDashScopeStream::test_stream_audio_final_block_declares_real_wav_sizes (tests/model_dashscope_test.py). They fail on unpatched main:

>       self.assertEqual(riff_size, 36 + len(pcm_full))
E       AssertionError: 4294967295 != 48
FAILED tests/model_openai_chat_test.py::TestOpenAIChatStream::test_stream_audio_final_block_declares_real_wav_sizes
FAILED tests/model_dashscope_test.py::TestDashScopeStream::test_stream_audio_final_block_declares_real_wav_sizes
============================== 2 failed in 10.11s ==============================

and pass with this change (2 passed); each test also parses the final block end-to-end with wave.open and asserts getnframes().

• Module tests: python3 -m pytest tests/model_openai_chat_test.py tests/model_dashscope_test.py → 39 passed.
• Full suite: python3 -m pytest tests (excluding the four files that fail to import without optional extras: blob_store_s3, mem0_agentscope_adapter, tui, webrtc_audio_transport) → 2457 passed, 159 failed; the 159 failures are byte-identical to the same run on unpatched main (they require external services: MySQL, Qdrant, MongoDB, Ollama, DingTalk).
• pre-commit run --all-files → all hooks Passed (mypy, black, flake8, pylint included).

Checklist

• An issue has been created for this PR
• I have read the CONTRIBUTING.md
• Docstrings are in Google style
• N/A — internal artifact fix; the streamed block's container format is not described in the documentation repository
• Code is ready for review

Fixes #3245